A Comparison of Unsupervised Classifiers on BATSE Catalog Data

摘要

We classify BATSE gamma-ray bursts using unsupervised clustering algorithms in order to compare classification with statistical clustering techniques. BATSE bursts detected with homogeneous trigger criteria and measured with a limited attribute set (duration, hardness, and fluence) are classified using four unsupervised algorithms (the concept hierarchy classifier ESX, the EM algorithm, the Kmeans algorithm, and a kohonen neural network). The classifiers prefer three-class solutions to two-class and four-class solutions. When forced to find two classes, the classifiers do not find the traditional long and short classes; many short soft events are placed in a class with the short hard bursts. When three classes are found, the classifiers clearly identify the short bursts, but place far more members in an intermediate duration soft class than have been found using statistical clustering techniques. It appears that the boundary between short faint and long bright bursts is more important to the classifiers than is the boundary between short hard and long soft bursts. We conclude that the boundary between short faint and long hard bursts is the result of data bias and poor attribute selection. We recommend that future gamma-ray burst classification avoid using extrinsic parameters such as fluence, and should instead concentrate on intrinsic properties such as spectral, temporal, and (when available) luminosity characteristics. Future classification should also be wary of correlated attributes (such as fluence and duration), as these bias classification results.